BLENDER: Enabling Local Search with a Hybrid Differential Privacy Model
summary
Presents a concrete algorithm, Blender, that mixes a small opt-in trusted-curator cohort with local-privacy clients to produce head lists of popular queries and click URLs. Uses Laplace noise and a two-stage k-ary randomized response, computes variance-aware weighted blending, and outputs unbiased probability estimates. On AOL and Yandex logs, using ~2–5% opt-in users yields NDCG >95% for head lists of 10–500; the system is differentially private and empirically restores utility lost by pure local methods.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:40Local Search
- 3:45Trustworthy Curator Model
- 5:21Differential Privacy Model
- 6:30Why Hybrid
- 8:37Blender
- 10:33Optin
- 14:51Measuring Utility
- 17:30Experimental Results
- 22:18Conclusion
- 23:57Questions
- 25:16Discussion
related talks
20mShare data without revealing personal information
58mLarge-Scale Private Learning, Part III
32mYou've got trust issues, we've got solutions: Differential Privacy
33mPractical Data Privacy
01h 43mPerspective
10mCommunicating Differential Privacy Guarantees to Data Subjects
27mYou've got trust issues, we've got solutions: Differential Privacy
01h 14mStatistical Inference and Privacy, Part III